|
From: | fotios |
Subject: | Re: add optimization function -> fmincon |
Date: | Fri, 20 Jan 2012 20:08:12 +0200 |
User-agent: | Mozilla/5.0 (X11; Linux i686; rv:8.0) Gecko/20111124 Thunderbird/8.0 |
On Fri, Jan 20, 2012 at 08:11:53AM -0500, Ben Abbott wrote:
On the IRC yesterday, it was suggested that ML's fmincon was sqp in disguise. Which algorithm of fmincon? Does it call the objective function with
infeasible parameters _during_ optimization, as sqp does?
... opinion on the idea of wrapping an m-file around sqp to provide an fmincon.m that is compatible with Matlab's? If a good generic optimization interface was designed, it could be useful to wrap sqp. Is fmincon the interface we want? If not, it could distract users focus from a possibly better interface. An issue which occurs to me: the fmincon interface does not allow computation of gradient without repeating the computation of the objective function at the same time (as far as I remember); this might not be efficient for all algorithms. Olaf As far as i remember, Olaf is right for ML's fmincon, though this could be fixed with the non-intelligent way: a conditional statement and a flag in the function that evaluates the objective function values and the gradient values and an anonymous function. I find a common interface a nice idea (i do not really care if it has the same interface as fmincon but for now lets say it has for those that do care to be satisfied). The one that will deal with the problem should certainly have a look at http://ab-initio.mit.edu/wiki/index.php/NLopt /Fotis |
[Prev in Thread] | Current Thread | [Next in Thread] |